Reducing EV range anxiety: How a simple AI model predicts port availability
November 21, 2025
Kostas Kollias, Research Scientist, Google Research
—
We developed a unique model to predict the probability that an EV charging port will be available at a given station within a certain number of minutes from the current time. This helps EV drivers plan their trips efficiently and minimize waiting time at charging stations.
The global transition to electric vehicles (EVs) is accelerating, creating a critical need for a reliable and efficient charging infrastructure. Beyond building more charging stations, maximizing their efficiency and minimizing “range anxiety”—the fear of running out of battery before reaching a destination or an available charger—is essential. This drove us to design an approach for EV routing that integrates charging stations based on battery level and destination.
Recently, we announced a lightweight, highly efficient prediction model addressing the question, “What is the probability that an EV charging port will be available at a specific station a number of minutes from now?” Our simple linear regression approach, co-designed with deployment infrastructure, is powerful yet relies on easily accessible features. This combination of real-world logic and machine learning delivers significant operational and user experience benefits.
### Creating the model
Our aim was to maximize predictive power while minimizing the feature set to ensure speed and low latency. After testing options like decision trees and simple neural networks, a linear regression model proved the most effective for this task. The model was trained using real-time data from charging networks to calculate port availability over time. Data came from two regions (California and Germany), with larger stations weighted more heavily due to their higher traffic.
### Features
The model uses the hour of the day as its main feature, treating each hour separately—for example, 9 AM and 5 PM are different inputs.
### Weights
The model learns specific weights for each hour representing the average rate at which ports become occupied or freed:
– Positive weight: ports tend to become occupied (e.g., 7 AM)
– Negative weight: ports tend to become free (e.g., 5 PM)
– Near-zero weight: little change (e.g., 3 AM)
These weights quantify the rate of change of charging port occupancy, reflecting predictable daily patterns tied to driver schedules.
The model only diverges from the current state when changes are significant (e.g., rush hours) or at large stations, which matches intuitive expectations for when updated predictions are useful.
### Experiments
We designed a rigorous evaluation focusing on 30- and 60-minute horizons. We tested 100 randomly selected stations, sampling every 30 minutes for a week. The model was benchmarked against a strong baseline: assuming port availability remains unchanged over the horizon. This simple baseline is tough to beat because most ports change state infrequently.
We measured accuracy using mean squared error (MSE) and mean absolute error (MAE), focusing especially on whether predictions correctly detect if at least one port is available.
### Results
Our linear regression model outperformed the “Keep Current State” baseline by effectively identifying rare but critical high-occupancy turnover moments. Tests focused on stations with 6+ ports revealed substantial accuracy improvements during peak times (8 AM and 8 PM).
Deploying this model reduced bad predictions by about 20% in morning peaks and 40% in evening peaks.
### Regional differences
While the shape of port occupancy change patterns was similar across regions, magnitudes varied enough that separate models for California and Germany performed better than pooled models. This highlights the importance of capturing regional usage patterns.
### Conclusion
We developed and deployed a lightweight linear regression model that efficiently predicts EV charging port availability. By focusing on simplicity, speed, and co-design with infrastructure, we bypassed the complexity and latency of more detailed models.
This approach provides a clear predictive advantage, reducing range anxiety and enabling smarter routing, ultimately supporting the continued adoption of electric mobility. Future work will explore extending prediction horizons to better support long-distance travel planning.
### Acknowledgements
We thank our collaborators Achir Ramadhan, Sreenivas Gollapudi, Shubham Gupta, Ilya Eyzerman, and Ivan Kuznetsov.
—
**Labels:** Algorithms & Theory, Product
—
**Other posts of interest:**
– [Real-time speech-to-speech translation (November 19, 2025)](https://research.google/blog/real-time-speech-to-speech-translation/)
– [Generative UI: A rich, custom, visual interactive user experience for any prompt (November 18, 2025)](https://research.google/blog/generative-ui-a-rich-custom-visual-interactive-user-experience-for-any-prompt/)
– [A new quantum toolkit for optimization (November 13, 2025)](https://research.google/blog/a-new-quantum-toolkit-for-optimization/)
